Treatment peptide prediction method based on nearest neighbor classifier

Abstract

The invention discloses a therapeutic peptide prediction method based on a nearest neighbor classifier, and belongs to the field of biological information. According to the method, peptide sequence features are extracted from QSP740 and CPP400 data sets by using a UniRep model, a kernel risk sensitive loss function is fused on the basis of a K-nearest neighbor algorithm to construct an objective function, and a multi-Laplacian matrix is constructed in combination with three similarity matrixes of an RBF function, cosine similarity and a Pearson's correlation coefficient to enhance model robustness. And finally, classifying the peptide sequence to be predicted by minimizing the objective function and evaluating the performance. According to the method, features are automatically extracted through deep learning, the model robustness is improved through a kernel risk loss function, and the uncertainty of data is processed by means of multi-view Laplace regularization, so that compared with an existing method, the treatment peptide type can be more accurately predicted, and the classification precision and the anti-noise capability can be improved.

Description

Technical Field

[0001] This invention relates to the field of bioinformatics, and in particular to a method for predicting therapeutic peptides based on a nearest neighbor classifier. Background Technology

[0002] Peptides are organic compounds composed of two or more amino acids and are irreplaceable building blocks for active substances such as enzymes, hormones, antibodies, and neurotransmitters in the human body. Compared to small-molecule amino acids and large-molecule proteins, peptides can be directly absorbed and utilized by the human body. They are important substances involved in various cellular functions, can synthesize cells, and regulate cellular functional activities, thereby comprehensively regulating human physiological functions, enhancing and exerting human physiological activity, and possessing important biological functions. Therapeutic peptides are a unique class of pharmaceutical preparations composed of a series of well-ordered amino acids. Therapeutic peptides have low immunogenicity and low production costs. Due to their unique biochemical properties and therapeutic potential, they are formulated into peptide drugs for the treatment of various diseases. Cell-penetrating peptides (CPPs), often referred to as Trojan horses, can naturally cross the lipid bilayer membrane protecting cells to deliver drugs and enhance the stimulation of immune cells. Quorum sensing peptides (QSPs) are important signaling molecules that drive quorum sensing, helping to establish intercellular communication, coordinate gene expression, and regulate physiological activities. QSPs have immunological importance as new diagnostic drugs and treatment methods in pathology. Therapeutic peptides have attracted much attention in recent years as potential drug candidates and delivery carriers. Scientists have attempted to efficiently and reliably produce synthetic peptides on a large scale. However, using biochemical experiments to verify peptide properties and predict peptide types is time-consuming and laborious. As an auxiliary tool to biochemical experiments, computational methods can improve the efficiency of therapeutic peptide detection.

[0003] Machine learning, as an effective and efficient computational method, has been applied to select and predict therapeutic peptides. Typically, peptide sequences are first converted into numerical features, and then machine learning algorithms are used to learn patterns from these features. QSP predictors, developed using powerful support vector machines and informative physicochemical properties, can effectively distinguish between QSP and non-QSP peptides. However, these predictors are prone to overfitting and have not been evaluated on independent datasets, making it impossible to determine their predictive ability for unknown peptides. Summary of the Invention

[0004] The purpose of this invention is to overcome one or more shortcomings of the prior art and provide a method for predicting therapeutic peptides based on a nearest neighbor classifier.

[0005] The objective of this invention is achieved through the following technical solution:

[0006] A method for predicting therapeutic peptides based on a nearest neighbor classifier includes the following steps:

[0007] S1. Feature extraction was performed on the QSP740 and CPP400 datasets, and feature vectors of peptide sequences were generated using the UniRep model;

[0008] S2. Construct an objective function by integrating the kernel risk-sensitive mean p-power error (KRP) loss function into the K nearest neighbor algorithm to measure the similarity between therapeutic peptide sequences;

[0009] S3. Construct a multi-Laplace matrix by combining the RBF function, cosine similarity, and Pearson correlation coefficient, and enhance the robustness of the model through Laplace regularization.

[0010] S4. Minimize the objective function to classify the peptide sequences to be classified into the category with the smallest function value, and evaluate the model performance using accuracy, sensitivity, specificity, and Matthew correlation coefficient.

[0011] Furthermore, in the feature extraction step of S1, the UniRep model uses a Long Short-Term Memory (LSTM) network to represent each amino acid in the protein sequence as a vector through one-hot encoding. After LSTM processing, an average pooling operation is performed to generate a 1900-dimensional peptide sequence feature vector.

[0012] Furthermore, in S1, the expression for the objective function is:

[0013] ;

[0014] Where x is a d-dimensional column vector, which is a 1900-dimensional feature vector generated by the UniRep model. Let j be the j-th feature, and c be the c-th category; It is the regularization coefficient of the penalty term in the weight vector. It is a hyperparameter that adjusts the degree of penalty imposed on the distance between the feature values ​​of the sample to be classified and the feature values ​​of its K nearest neighbors. It is a Gaussian function. is the variance of the Gaussian function, and p is a configurable power value. The d-dimensional feature matrix of the K nearest neighbors of the test sample in class c. Let K be the feature row vectors of the K nearest neighbors in the j-th dimension; Let K be the weight vectors of the nearest neighbors. Let f be the weight of the f-th nearest neighbor; It is the regularization coefficient of the Laplace regularization term. Let f be an adjacency matrix, and let g be a matrix where f is a K-nearest neighbor of g. The similarity is calculated using a similarity matrix; otherwise, it is 0.

[0015] Furthermore, the objective function is transformed into a form containing a normalized Laplace matrix:

[0016] ;

[0017] Among them, the normalized Laplace matrix satisfy:

[0018] ;

[0019] Where D is a diagonal matrix. That is, the i-th element is equal to the sum of the elements in the i-th row of matrix W. It is the i-th sample point.

[0020] Furthermore, in S3, the fusion expression for the dolaplace matrix is:

[0021] ;

[0022] in, The merged dolaplace matrix, Let be the Laplace matrix of the v-th normalization, and m be the number of different methods for calculating the Laplace matrix. Let v be the weights of the normalized Laplace matrix. These are the configurable power values ​​corresponding to the weights of the Laplace matrix.

[0023] Furthermore, the objective function of the dolaplace matrix with the dolaplace regularization term is:

[0024] .

[0025] Furthermore, in step S4, the weight vector is obtained by minimizing the objective function. The expression is:

[0026] ;

[0027] in, , , is the variance of the Gaussian function, and d is the total dimension of the features. It is the regularization coefficient of the penalty term in the weight vector. is the regularization coefficient of the Laplace regularization term, and X is the feature matrix of the samples to be classified. It is a K-order identity matrix. It is the transpose of the d-dimensional feature matrix of the K nearest neighbors of the test sample in class c. It is a diagonal matrix with the following elements:

[0028] ;

[0029] in, Let be the feature difference between the sample to be classified and its K nearest neighbors in class c, given the d-dimensional feature matrix.

[0030] Furthermore, in S2, the kernel risk-sensitive loss function maps the data to a high-dimensional space using a kernel method, utilizes p-power error to reduce the impact of noise and outliers on the model, and the loss function has clear upper and lower bounds to avoid function value explosion.

[0031] Furthermore, in S3, the Laplacian regularization term is represented by the Laplacian matrix combined with the Laplacian mapping, which is used to maintain the local structure of the data and prevent the model from overfitting, thereby improving the model's generalization ability.

[0032] Furthermore, in some embodiments, multi-view learning is used to estimate the score of each similarity and assign corresponding weights, and the uncertainty in the data is handled by integrating the weighted normalization results of multiple similarity matrices.

[0033] The beneficial effects of this invention are:

[0034] (1) Using the deep learning model UniRep to extract peptide sequence features, on the one hand, the characteristics of peptide sequences are automatically learned, and on the other hand, numerical features can be obtained without further processing of the model output features, which can be directly used to calculate the objective function;

[0035] (2) By introducing a kernel risk-sensitive loss function into the k-nearest neighbor model, the function maps the data to a higher-dimensional space through the kernel method, making the data more separable in the higher-dimensional space. At the same time, thanks to the clear upper and lower bounds of the loss function, the big explosion of function values ​​is avoided, thereby improving the robustness and classification accuracy of the model.

[0036] (3) Based on the preservation of local structure, a multi-view Laplace regularization method is proposed. By combining multiple similarity matrices and performing weighted normalization, it can better handle the uncertainty and noise in the data. Attached Figure Description

[0037] Figure 1 This is a flowchart of a therapeutic peptide prediction method based on a nearest neighbor classifier.

[0038] Figure 2 A flowchart illustrating the algorithm construction steps for a therapeutic peptide prediction method based on nearest neighbor classifier;

[0039] Figure 3 The complete flowchart of the therapeutic peptide prediction method based on the nearest neighbor classifier provided in the embodiment is shown. Detailed Implementation

[0040] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0041] Example 1

[0042] See Figure 1 This paper provides a method for predicting therapeutic peptides based on a nearest neighbor classifier, comprising the following steps:

[0043] S1. Feature extraction was performed on the QSP740 and CPP400 datasets, and feature vectors of peptide sequences were generated using the UniRep model;

[0044] S2. Construct an objective function by integrating the kernel risk-sensitive mean p-power error (KRP) loss function into the K nearest neighbor algorithm to measure the similarity between therapeutic peptide sequences;

[0045] S3. Construct a multi-Laplace matrix by combining the RBF function, cosine similarity, and Pearson correlation coefficient, and enhance the robustness of the model through Laplace regularization.

[0046] S4. Minimize the objective function to classify the peptide sequences to be classified into the category with the smallest function value, and evaluate the model performance using accuracy, sensitivity, specificity, and Matthew correlation coefficient.

[0047] Furthermore, in the feature extraction step of S1, the UniRep model uses a Long Short-Term Memory (LSTM) network to represent each amino acid in the protein sequence as a vector through one-hot encoding. After LSTM processing, an average pooling operation is performed to generate a 1900-dimensional peptide sequence feature vector.

[0048] Furthermore, in S1, the expression for the objective function is:

[0049] ;

[0050] Where x is a d-dimensional column vector, which is a 1900-dimensional feature vector generated by the UniRep model. Let j be the j-th feature, and c be the c-th category; It is the regularization coefficient of the penalty term in the weight vector. It is a hyperparameter that adjusts the degree of penalty imposed on the distance between the feature values ​​of the sample to be classified and the feature values ​​of its K nearest neighbors. It is a Gaussian function. is the variance of the Gaussian function, and p is a configurable power value. The d-dimensional feature matrix of the K nearest neighbors of the test sample in class c. Let K be the feature row vectors of the K nearest neighbors in the j-th dimension; Let K be the weight vectors of the nearest neighbors. Let f be the weight of the f-th nearest neighbor; It is the regularization coefficient of the Laplace regularization term. Let f be an adjacency matrix, and let g be a matrix where f is a K-nearest neighbor of g. The similarity is calculated using a similarity matrix; otherwise, it is 0.

[0051] Furthermore, the objective function is transformed into a form containing a normalized Laplace matrix:

[0052] ;

[0053] Among them, the normalized Laplace matrix satisfy:

[0054] ;

[0055] Where D is a diagonal matrix. That is, the i-th element is equal to the sum of the elements in the i-th row of matrix W. It is the i-th sample point.

[0056] Furthermore, in S3, the fusion expression for the dolaplace matrix is:

[0057] ;

[0058] in, The merged dolaplace matrix, Let be the Laplace matrix of the v-th normalization, and m be the number of different methods for calculating the Laplace matrix. Let v be the weights of the normalized Laplace matrix. These are the configurable power values ​​corresponding to the weights of the Laplace matrix.

[0059] Furthermore, the objective function of the dolaplace matrix with the dolaplace regularization term is:

[0060] .

[0061] Furthermore, in step S4, the weight vector is obtained by minimizing the objective function. The expression is:

[0062] ;

[0063] in, , , is the variance of the Gaussian function, and d is the total dimension of the features. It is the regularization coefficient of the penalty term in the weight vector. is the regularization coefficient of the Laplace regularization term, and X is the feature matrix of the samples to be classified. It is a K-order identity matrix. It is the transpose of the d-dimensional feature matrix of the K nearest neighbors of the test sample in class c. It is a diagonal matrix with the following elements:

[0064] ;

[0065] in, Let be the feature difference between the sample to be classified and its K nearest neighbors in class c, given the d-dimensional feature matrix.

[0066] Furthermore, in S2, the kernel risk-sensitive loss function maps the data to a high-dimensional space using a kernel method, utilizes p-power error to reduce the impact of noise and outliers on the model, and the loss function has clear upper and lower bounds to avoid function value explosion.

[0067] Furthermore, in S3, the Laplacian regularization term is represented by the Laplacian matrix combined with the Laplacian mapping, which is used to maintain the local structure of the data and prevent the model from overfitting, thereby improving the model's generalization ability.

[0068] Furthermore, in some embodiments, multi-view learning is used to estimate the score of each similarity and assign corresponding weights, and the uncertainty in the data is handled by integrating the weighted normalization results of multiple similarity matrices.

[0069] Example 2: Complete implementation process of the therapeutic peptide prediction method based on kernel risk nearest neighbor classifier:

[0070] Specific implementation of data preparation and feature extraction: Dataset acquisition and preprocessing: The QSP740 and CPP400 datasets were obtained from publicly available bioinformatics databases. QSP740 contains experimentally validated quorum sensing peptide sequences, and CPP400 contains cell-penetrating peptide sequences. For example... Figure 2 The "Extract Sequence Features" module first cleans the dataset, removing repetitive sequences and peptides of abnormal length, and then divides the data into training and testing sets in an 8:2 ratio. The training set is used for model parameter learning, while the testing set is used to independently evaluate model performance. For each peptide sequence, one-hot encoding is used to convert amino acids into numerical vectors—20 standard amino acids correspond to 20-dimensional vectors, and a 1 value at a specific position in the vector represents the corresponding amino acid, ensuring the completeness of the sequence information conversion.

[0071] Feature extraction process of the UniRep model: Feature extraction is performed using the UniRep model based on a Long Short-Term Memory (LSTM) network. For example... Figure 2 As shown, this model processes amino acid sequence vectors through a recursive structure: the one-hot encoded amino acid vectors are sequentially input into an LSTM network, and a gating mechanism is used to capture long-range dependencies in the sequence (such as spatial interactions or functional correlations between amino acids). After sequence processing, average pooling is used to integrate the outputs of each time step of the LSTM into a fixed-dimensional feature vector. Specifically, the UniRep model automatically learns patterns in the sequence, directly generating a 1900-dimensional peptide sequence feature vector without the need for manual design of physicochemical properties and other features. This vector contains deep semantic information of the sequence and can effectively characterize the structural and functional properties of the peptide.

[0072] Objective function construction and implementation of regularization methods:

[0073] Fusion of the K-nearest neighbor framework and the kernel risk loss function: such as Figure 2 As shown in the "Constructing the Objective Function" module, for each test sample, the K nearest neighbor samples are first searched in each category of the training set using a distance metric (such as Euclidean distance) to form a feature matrix V_c. The adjacency matrix W_c is constructed according to the following rule: when two nearest neighbor samples are each other's K nearest neighbors, the weights at corresponding positions are calculated using a similarity matrix (such as cosine similarity); otherwise, they are set to 0. Based on this, the objective function is constructed by fusing the kernel risk-sensitive mean p-power error (KRP) loss function: the kernel method maps the data to a high-dimensional space, making originally linearly inseparable peptide sequences easier to distinguish in the high-dimensional space; the p-power error structure is used to reduce the impact of noise and outliers on the model—when there is a bias in the sample features, the nonlinearity of the p-power function can suppress the weights of outliers, ensuring the robustness of the model to noise.

[0074] Construction and regularization of the Dolaplace matrix:

[0075] like Figure 2 The "Constructing Regularization Methods" module and Figure 3 The "Calculate Multiple Weight Matrices" node shows three different methods for calculating similarity matrices to comprehensively characterize the similarity between peptide sequences:

[0076] RBF function: Measures the similarity between feature vectors using a Gaussian kernel function, adapting to non-linearly distributed data;

[0077] Cosine similarity: measures the directional consistency of feature vectors and is suitable for characterizing the similarity of sequence patterns;

[0078] Pearson correlation coefficient: assesses the linear correlation of feature vectors and captures the co-variation patterns of sequence features. After generating a corresponding normalized Laplacian matrix for each similarity matrix, a multi-view learning method is used to weight and fuse them.

[0079] After generating a corresponding normalized Laplacian matrix for each similarity matrix, a multi-view learning method is used to weight and fuse them. The weight allocation satisfies the normalization condition ( The fused multi-Laplacian matrix is ​​embedded into the objective function through a Laplacian regularization term to maintain the local geometry of the data—that is, similar peptide sequences maintain high correlation in the feature space, prevent model overfitting, and improve generalization ability.

[0080] Classification decision-making process and performance evaluation system:

[0081] Weight vector calculation and classification decision implementation:

[0082] like Figure 3 As shown in the "Complete Algorithm Flow", the weight vectors of the K nearest neighbors are solved by minimizing the objective function. This process transforms the objective function into a matrix operation problem, obtaining the optimal weights by solving a system of linear equations. Specifically, the feature vector of the test sample is combined with the nearest neighbor feature matrix, and the weight vector is obtained using matrix inversion. The nearest neighbor sample with a larger weight has a greater impact on the classification decision. Finally, the test sample is classified into the category with the smallest objective function value—that is, the sample has the highest similarity to the nearest neighbor sample of the corresponding category, and the highest probability of belonging to that category.

[0083] The model was validated in multiple dimensions using industry-standard evaluation metrics:

[0084] Accuracy: Measures the proportion of samples correctly classified by the model, reflecting the overall classification ability;

[0085] Sensitivity: Evaluates the model's ability to identify positive samples (such as QSPs);

[0086] Specificity: Measures the model's ability to distinguish negative samples (such as non-QSP samples);

[0087] Matthew correlation coefficient: A comprehensive indicator that balances true positives, true negatives, false positives, and false negatives; suitable for imbalanced datasets. For example... Figure 3 As shown, evaluation is carried out through 10-fold cross-validation: the dataset is repeatedly divided 10 times, with 90% of the data used for training and 10% for testing each time, and the average index is taken as the final result to ensure the reliability and generalization of the evaluation results.

[0088] like Figure 3As shown, evaluation is carried out through 10-fold cross-validation: the dataset is repeatedly divided 10 times, with 90% of the data used for training and 10% for testing each time, and the average index is taken as the final result to ensure the reliability and generalization of the evaluation results.

[0089] Compared to traditional feature extraction methods (such as statistical analysis of physicochemical properties), the UniRep model automatically captures hidden patterns in sequences through deep learning, avoiding the subjectivity and limitations of manual feature design. For example, for peptide sequences with similar functions, the feature vectors generated by UniRep are closer in high-dimensional space, while traditional methods may lead to inaccurate feature representation due to ignoring sequence context.

[0090] The kernel risk loss function maps data to a high-dimensional space through kernel mapping, transforming linearly inseparable problems into linearly separable ones. Simultaneously, it utilizes the boundedness of p-power errors to prevent function value explosion. When outliers are present, the growth rate of the loss function slows down as the error increases, thus reducing the impact of outliers on model parameters. This characteristic is particularly important in biological data processing—experimentally obtained peptide sequence features often contain noise due to measurement errors.

[0091] like Figure 3 As shown in the "Laplacian Mapping" node, multi-Laplacian regularization ensures that the data representation learned by the model retains the local structure of the original space by constraining the weight differences of similar samples. For example, if two QSP sequences are similar in the original feature space, their corresponding weight vector components will have small differences, thus being assigned similar decision weights during classification and avoiding classification bias caused by feature perturbations.

[0092] Example 3: Multi-view Laplace regularization-based optimization method for predicting therapeutic peptides:

[0093] Dynamic optimization process of similarity matrix weights: Construction of a multi-view learning framework: such as Figure 2 The "Constructing Regularization Methods" module and Figure 3 As shown in the "Matrix Fusion" node, the core of the multi-view learning framework lies in integrating the complementary information of three similarity matrices. Each similarity matrix measures sequence similarity from a different perspective: the RBF function is sensitive to feature distance, cosine similarity focuses on directional consistency, and the Pearson correlation coefficient emphasizes linear correlation. To fully utilize this information, cross-validation is used to search for the optimal weight combination. Where v=1,2,3 correspond to three similarity matrices, and the weights satisfy the following: and .

[0094] The specific implementation steps of weight optimization are as follows:

[0095] Search space definition: Using grid search or random search methods, define the candidate range for the weight parameters, for example... , , and This ensures that the weighting of each similarity matrix is ​​reasonable;

[0096] Evaluation index setting: The Matthew correlation coefficient with 10-fold cross-validation is used as the optimization objective. The candidate weight combinations are traversed, and the weight that makes the evaluation index optimal is selected as the final fusion weight.

[0097] Constraint handling: The weight normalization constraint is incorporated into the optimization process using the Lagrange multiplier method to ensure the physical meaning of the weight allocation—the larger the weight of the similarity matrix, the greater the contribution of its corresponding Laplacian matrix in the fusion, thus adapting to the similarity measurement needs under different data distributions.

[0098] Joint optimization implementation of objective function and regularization term:

[0099] Construction of the objective function with weighted optimization:

[0100] like Figure 3 The "Calculate Objective Function" node shows the weights of the similarity matrix. As an optimization variable, the nearest neighbor weight vector Simultaneous solution. The optimization objective function is expressed as:

[0101] ;

[0102] Among them, the dolaplace matrix , The weight exponent is used to adjust the non-linear distribution of the weights. This objective function simultaneously considers the model's fitting ability (KRP loss term), regularization constraints (Laplacian regularization term), and the weight allocation of the similarity matrix, forming an end-to-end optimization framework.

[0103] Weighting index Controlling the non-linearity of the similarity matrix weights: when When the value is greater than 1, the weight difference is amplified, making the contribution of the advantage similarity matrix more prominent; when... When the weights are equal to 1, the weights are linearly fused. In implementation, this can be adjusted according to the data distribution characteristics. For example, for data with significant differences in the similarity matrix, increase To enhance the role of the dominance matrix. Laplace regularization term. Essentially, it measures the weight vector. Smoothness on graph structures defined by the dolaplace matrix. The value of this regularization term is small when the weight differences between similar samples are small, thus constraining the model to learn a smooth weight distribution and avoiding overfitting. This mechanism is consistent with the smoothness assumption in graph semi-supervised learning, ensuring the model's generalization ability on unseen samples.

[0104] The iterative process for optimizing the solution:

[0105] Initialization phase: Randomly initialize the weights of the similarity matrix , which satisfies the normalization condition;

[0106] Weight update: Fixed Update using gradient descent or grid search. Minimize the objective function;

[0107] Vector update: fixed Solve for the nearest neighbor weight vector The closed-form solution is obtained by matrix inversion.

[0108] Iterative convergence: Repeat the above steps until the objective function value converges, or the preset number of iterations is reached, to ensure convergence. and At the same time, it is in the optimal state.

[0109] like Figure 3 The "Predict Labels for Samples to be Classified" node demonstrates how to apply the optimized model to a test set that was not used in training, comparing the performance differences of different weighting strategies: weights set empirically (such as equal-weight fusion) have low classification accuracy for unknown samples because they are not adapted to the data distribution; weights optimized through multi-view learning can automatically adjust the contribution of the similarity matrix according to data features, significantly improving classification accuracy. For example, when peptide sequences in the test set have a strong non-linear distribution, the weights of the RBF similarity matrix will automatically increase, strengthening the model's ability to identify non-linear patterns.

[0110] Experimental verification of noise resistance performance involved artificially adding noise (such as randomly perturbing part of the feature vector) to the test set and comparing the classification stability of the model in noisy environments. The kernel risk loss function, through the p-power error structure, ensures that the loss value of noisy samples grows at a lower rate than that of normal samples, thereby reducing their impact on classification decisions. The multi-Laplace regularized fusion matrix maintains the local structure of the data, reducing the difference between the weight vector of noisy samples and the weight vector of similar normal samples, thus avoiding classification errors caused by feature perturbation. The synergistic effect of multi-view fusion filters noise from different angles through three similarity matrices: the RBF function is insensitive to local noise, cosine similarity focuses on the overall pattern, and Pearson correlation coefficient suppresses abnormal fluctuations in feature dimensions. The weighted fusion of the three forms a multi-layered noise resistance mechanism.

[0111] In real-world scenarios, multi-view learning addresses the limitations of single similarity measures by integrating multiple similarity matrices: a single RBF similarity measure may fail due to the curse of dimensionality in the feature space; a single cosine similarity measure cannot capture the absolute differences in feature dimensions.

[0112] A single Pearson correlation coefficient is insufficient for handling non-linear correlation patterns in sequences. Multi-view fusion, through dynamic weight adjustment, allows the model to adaptively select the most suitable similarity metric for the current data, thus improving classification robustness.

[0113] In practical applications, for newly discovered peptide sequences, this method can accurately predict whether they are therapeutic peptides (such as QSP or CPP) through deep feature extraction and multi-view similarity measurement, providing computational support for experimental verification. In the process of peptide drug design, this model can be used to evaluate the similarity between candidate peptide sequences and known therapeutic peptides, guiding the rational design of highly active peptides. When faced with massive peptide sequence libraries (such as peptide libraries synthesized by combinatorial chemistry), this method can efficiently screen peptides with potential therapeutic functions, reducing experimental screening costs.

[0114] The above description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the concept described herein through the above teachings or related technologies or knowledge. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. A method for predicting therapeutic peptides based on a nearest neighbor classifier, characterized in that, Includes the following steps: S1. Feature extraction was performed on the QSP740 and CPP400 datasets, and feature vectors of peptide sequences were generated using the UniRep model; S2. Construct an objective function that integrates a kernel risk-sensitive mean p-power error loss function with the K nearest neighbor algorithm to measure the similarity between therapeutic peptide sequences; S3. Construct a multi-Laplace matrix by combining the RBF function, cosine similarity, and Pearson correlation coefficient, and enhance the robustness of the model through Laplace regularization. S4. Minimize the objective function to classify the peptide sequences to be classified into the category with the smallest function value, and evaluate the model performance using accuracy, sensitivity, specificity, and Matthew correlation coefficient.

2. The method according to claim 1, characterized in that, In the feature extraction step of S1, the UniRep model uses a long short-term memory network to represent each amino acid in the protein sequence as a vector through one-hot encoding. After LSTM processing, an average pooling operation is performed to generate a 1900-dimensional peptide sequence feature vector.

3. The method according to claim 1, characterized in that, In S1, the expression for the objective function is: ； Where x is a d-dimensional column vector, which is a 1900-dimensional feature vector generated by the UniRep model. Let j be the j-th feature, and c be the c-th category; It is the regularization coefficient of the penalty term in the weight vector. It is a hyperparameter that adjusts the degree of penalty imposed on the distance between the feature values ​​of the sample to be classified and the feature values ​​of its K nearest neighbors. It is a Gaussian function. is the variance of the Gaussian function, and p is a configurable power value. The d-dimensional feature matrix of the K nearest neighbors of the test sample in class c. Let K be the feature row vectors of the K nearest neighbors in the j-th dimension; Let K be the weight vectors of the nearest neighbors. Let f be the weight of the f-th nearest neighbor; It is the regularization coefficient of the Laplace regularization term. Let f be an adjacency matrix, and let g be a matrix where f is a K-nearest neighbor of g. The similarity is calculated using a similarity matrix; otherwise, it is 0.

4. The method according to claim 3, characterized in that, The objective function is transformed into a form that includes a normalized Laplacian matrix: ； Among them, the normalized Laplace matrix satisfy: ； Where D is a diagonal matrix. That is, the i-th element is equal to the sum of the elements in the i-th row of matrix W. It is the i-th sample point.

5. The method according to claim 1, characterized in that, In S3, the fusion expression of the dolaplace matrix is: ； in, The merged dolaplace matrix, Let be the Laplace matrix of the v-th normalization, and m be the number of different methods for calculating the Laplace matrix. Let v be the weights of the normalized Laplace matrix. These are the configurable power values ​​corresponding to the weights of the Laplace matrix.

6. The method according to claim 5, characterized in that, The objective function of the dolaplace matrix with the dolaplace regularization term is: 。 7. The method according to claim 1, characterized in that, In step S4, the weight vector is obtained by minimizing the objective function. The expression is: ； in, , , is the variance of the Gaussian function, and d is the total dimension of the features. It is the regularization coefficient of the penalty term in the weight vector. is the regularization coefficient of the Laplace regularization term, and X is the feature matrix of the samples to be classified. It is a K-order identity matrix. It is the transpose of the d-dimensional feature matrix of the K nearest neighbors of the test sample in class c. It is a diagonal matrix with the following elements: ； in, Let be the feature difference between the sample to be classified and its K nearest neighbors in class c, given the d-dimensional feature matrix.

8. The method according to claim 1, characterized in that, In S2, the kernel risk-sensitive loss function maps the data to a high-dimensional space using a kernel method, and uses p-power error to reduce the impact of noise and outliers on the model.

9. The method according to claim 1, characterized in that, In S3, the Laplacian regularization term is represented by the Laplacian matrix combined with the Laplacian mapping, which is used to preserve the local structure of the data and prevent the model from overfitting.

10. The method according to any one of claims 1-9, characterized in that, Multi-view learning is used to estimate the score of each similarity and assign corresponding weights. The uncertainty in the data is handled by integrating the weighted normalization results of multiple similarity matrices.

Images